Fairness
How models come to treat groups of people differently, the group-fairness criteria that make "fair" precise (demographic parity, equalized odds, calibration), the theorems proving they conflict whenever base rates differ, individual fairness, the COMPAS debate as it was actually argued, and what mitigation can and cannot do.
50 min
After this lesson you can
- Name the main routes by which data and models produce unfair outcomes
- Define independence (demographic parity), separation (equalized odds) and sufficiency (calibration within groups) precisely
- Derive Chouldechova's identity and state exactly what Kleinberg, Mullainathan and Raghavan's theorem proves
- Explain individual fairness and why it needs a similarity metric
- Describe pre-, in- and post-processing mitigations and their costs
In this lesson
- How models become unfair
- Three definitions of group fairness
- Watching them collide
- The impossibility results
- Individual fairness
- COMPAS, as documented
- Mitigation: before, during and after training
This lesson is part of Descent Pro
Descent Pro unlocks every lesson on the route, from base camp to the summit. Every plan starts with a free trial.